linux_play.py

import os
import sys
import time
import json
import argparse
from pathlib import Path
from collections import OrderedDict

import cv2
import numpy as np
from PIL import Image

from nitrogen.shared import BUTTON_ACTION_TOKENS, PATH_REPO

# Do not import nitrogen.game_env at module import time.
# game_env imports pyautogui/dxcam, which fails on headless Linux or non-Windows.
try:
    from nitrogen.inference_viz import create_viz, VideoRecorder
    _VIZ_IMPORT_ERROR = None
except Exception as exc:  # Allows --no-video debugging even if video deps fail.
    create_viz = None
    VideoRecorder = None
    _VIZ_IMPORT_ERROR = exc


class _NoOpGamepad:
    def update(self):
        pass


class _NoOpGamepadEmulator:
    def __init__(self):
        self.gamepad = _NoOpGamepad()

    def press_button(self, button):
        pass

    def release_button(self, button):
        pass


class MockGamepadEnv:
    """Linux/debug replacement for GamepadEnv.

    It has the same small API surface used by this script:
    reset/pause/unpause/step/close and gamepad_emulator.
    step() returns a random RGB PIL image as obs.
    """

    def __init__(
        self,
        game,
        game_speed=1.0,
        env_fps=60,
        async_mode=True,
        width=1280,
        height=720,
        seed=0,
    ):
        self.game = game
        self.game_speed = game_speed
        self.env_fps = env_fps
        self.async_mode = async_mode
        self.width = width
        self.height = height
        self.rng = np.random.default_rng(seed)
        self.step_count = 0
        self.gamepad_emulator = _NoOpGamepadEmulator()

    def _random_obs(self):
        arr = self.rng.integers(
            0,
            256,
            size=(self.height, self.width, 3),
            dtype=np.uint8,
        )
        return Image.fromarray(arr, mode="RGB")

    def reset(self):
        self.step_count = 0
        return self._random_obs()

    def pause(self):
        pass

    def unpause(self):
        pass

    def close(self):
        pass

    def step(self, action):
        self.step_count += 1
        obs = self._random_obs()
        reward = 0.0
        terminated = False
        truncated = False
        info = {
            "mock_env": True,
            "step_count": self.step_count,
            "game": self.game,
        }
        return obs, reward, terminated, truncated, info


class VisualDebugGamepadEnv:
    """Linux/debug environment with meaningful RGB observations.

    This env does not need a display server. It renders a simple 2D scene with OpenCV:
    - a controllable player dot
    - a target dot
    - recent button states
    - left/right joystick values
    Therefore action outputs visibly change the next observation/video.
    """

    def __init__(
        self,
        game,
        game_speed=1.0,
        env_fps=60,
        async_mode=True,
        width=1280,
        height=720,
        seed=0,
    ):
        self.game = game
        self.game_speed = game_speed
        self.env_fps = env_fps
        self.async_mode = async_mode
        self.width = width
        self.height = height
        self.rng = np.random.default_rng(seed)
        self.step_count = 0
        self.gamepad_emulator = _NoOpGamepadEmulator()
        self.player = np.array([width * 0.25, height * 0.5], dtype=np.float32)
        self.velocity = np.zeros(2, dtype=np.float32)
        self.target = np.array([width * 0.78, height * 0.5], dtype=np.float32)
        self.last_action = make_zero_action()
        self.trail = []

    @staticmethod
    def _scalar(v):
        if isinstance(v, np.ndarray):
            return float(np.asarray(v).reshape(-1)[0])
        return float(v)

    def reset(self):
        self.step_count = 0
        self.player = np.array([self.width * 0.25, self.height * 0.5], dtype=np.float32)
        self.velocity[:] = 0
        self.target = np.array([
            self.rng.uniform(self.width * 0.55, self.width * 0.9),
            self.rng.uniform(self.height * 0.2, self.height * 0.8),
        ], dtype=np.float32)
        self.trail = []
        self.last_action = make_zero_action()
        return self._render()

    def pause(self):
        pass

    def unpause(self):
        pass

    def close(self):
        pass

    def step(self, action):
        self.step_count += 1
        self.last_action = action

        lx = self._scalar(action.get("AXIS_LEFTX", 0)) / 32767.0
        ly = self._scalar(action.get("AXIS_LEFTY", 0)) / 32767.0
        rx = self._scalar(action.get("AXIS_RIGHTX", 0)) / 32767.0
        rt = self._scalar(action.get("RIGHT_TRIGGER", 0)) / 255.0
        lt = self._scalar(action.get("LEFT_TRIGGER", 0)) / 255.0

        # D-pad also moves the player, making binary button mapping visible.
        if action.get("DPAD_LEFT", 0):
            lx -= 1.0
        if action.get("DPAD_RIGHT", 0):
            lx += 1.0
        if action.get("DPAD_UP", 0):
            ly -= 1.0
        if action.get("DPAD_DOWN", 0):
            ly += 1.0

        speed = 10.0 + 18.0 * max(rt, 0.0)
        damping = 0.78 if not action.get("SOUTH", 0) else 0.55
        accel = np.array([lx, ly], dtype=np.float32) * speed
        self.velocity = self.velocity * damping + accel
        if action.get("EAST", 0):
            # A visible jump/boost.
            self.velocity += np.array([18.0, -12.0], dtype=np.float32)
        if action.get("WEST", 0):
            self.velocity *= 0.35
        if action.get("NORTH", 0):
            # Move target so button effects are visible even without player motion.
            self.target = np.array([
                self.rng.uniform(self.width * 0.1, self.width * 0.9),
                self.rng.uniform(self.height * 0.15, self.height * 0.85),
            ], dtype=np.float32)

        self.player += self.velocity
        margin = 35
        self.player[0] = np.clip(self.player[0], margin, self.width - margin)
        self.player[1] = np.clip(self.player[1], margin, self.height - margin)

        self.trail.append(tuple(self.player.astype(int)))
        self.trail = self.trail[-80:]

        dist = float(np.linalg.norm(self.player - self.target))
        reward = -dist / max(self.width, self.height)
        if dist < 45:
            reward = 1.0
            self.target = np.array([
                self.rng.uniform(self.width * 0.1, self.width * 0.9),
                self.rng.uniform(self.height * 0.15, self.height * 0.85),
            ], dtype=np.float32)

        obs = self._render(lx=lx, ly=ly, rx=rx, lt=lt, rt=rt, reward=reward)
        terminated = False
        truncated = False
        info = {
            "debug_env": "visual",
            "step_count": self.step_count,
            "distance_to_target": dist,
            "game": self.game,
        }
        return obs, reward, terminated, truncated, info

    def _render(self, lx=0.0, ly=0.0, rx=0.0, lt=0.0, rt=0.0, reward=0.0):
        img = np.zeros((self.height, self.width, 3), dtype=np.uint8)
        img[:] = (28, 30, 36)

        # Background grid.
        for x in range(0, self.width, 80):
            cv2.line(img, (x, 0), (x, self.height), (45, 48, 56), 1)
        for y in range(0, self.height, 80):
            cv2.line(img, (0, y), (self.width, y), (45, 48, 56), 1)

        # A few fixed obstacles for spatial reference.
        cv2.rectangle(img, (self.width // 2 - 80, 130), (self.width // 2 + 80, 210), (80, 80, 95), -1)
        cv2.rectangle(img, (self.width // 2 - 120, self.height - 230), (self.width // 2 + 120, self.height - 150), (80, 80, 95), -1)

        # Trail.
        if len(self.trail) >= 2:
            for i in range(1, len(self.trail)):
                thickness = 1 + i // 25
                cv2.line(img, self.trail[i - 1], self.trail[i], (90, 180, 255), thickness)

        # Target and player.
        target_xy = tuple(self.target.astype(int))
        player_xy = tuple(self.player.astype(int))
        cv2.circle(img, target_xy, 34, (70, 220, 100), -1)
        cv2.circle(img, target_xy, 45, (70, 220, 100), 2)
        cv2.circle(img, player_xy, 28, (70, 130, 255), -1)
        cv2.circle(img, player_xy, 36, (230, 235, 245), 2)

        # Direction vector from joystick.
        tip = (int(self.player[0] + lx * 90), int(self.player[1] + ly * 90))
        cv2.arrowedLine(img, player_xy, tip, (255, 220, 120), 5, tipLength=0.3)

        # Text panel.
        panel_x, panel_y = 30, 35
        line_h = 32
        cv2.putText(img, "Linux VisualDebugGamepadEnv", (panel_x, panel_y), cv2.FONT_HERSHEY_SIMPLEX, 0.9, (245, 245, 245), 2)
        cv2.putText(img, f"step={self.step_count} reward={reward:.3f}", (panel_x, panel_y + line_h), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (230, 230, 230), 2)
        cv2.putText(img, f"left=({lx:+.2f},{ly:+.2f}) right_x={rx:+.2f} LT={lt:.2f} RT={rt:.2f}", (panel_x, panel_y + 2 * line_h), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (230, 230, 230), 2)

        pressed = [k for k, v in self.last_action.items() if not isinstance(v, np.ndarray) and bool(v)]
        pressed_text = "pressed: " + (", ".join(pressed[:10]) if pressed else "none")
        cv2.putText(img, pressed_text, (panel_x, panel_y + 3 * line_h), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (230, 230, 230), 2)

        # Button lamps.
        button_names = ["WEST", "SOUTH", "EAST", "NORTH", "START", "BACK"]
        bx = panel_x
        by = panel_y + 5 * line_h
        for idx, name in enumerate(button_names):
            on = bool(self.last_action.get(name, 0))
            color = (80, 220, 120) if on else (75, 75, 85)
            x0 = bx + idx * 125
            cv2.rectangle(img, (x0, by), (x0 + 105, by + 42), color, -1)
            cv2.putText(img, name, (x0 + 8, by + 28), cv2.FONT_HERSHEY_SIMPLEX, 0.55, (10, 10, 10) if on else (210, 210, 210), 2)

        return Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))


class GymRGBGamepadEnv:
    """Optional Gymnasium RGB-array environment for Linux.

    Example:
        pip install "gymnasium[classic-control]" pygame
        python play_linux_debug_visual.py --debug-env gym --gym-id CartPole-v1 --mock-policy --max-steps 10
    """

    def __init__(self, game, game_speed=1.0, env_fps=60, async_mode=True, gym_id="CartPole-v1", seed=0, **kwargs):
        os.environ.setdefault("SDL_VIDEODRIVER", "dummy")
        import gymnasium as gym

        self.game = game
        self.env_fps = env_fps
        self.gym_id = gym_id
        self.seed = seed
        self.step_count = 0
        self.gamepad_emulator = _NoOpGamepadEmulator()
        self.env = gym.make(gym_id, render_mode="rgb_array")
        self.discrete_n = getattr(self.env.action_space, "n", None)

    @staticmethod
    def _scalar(v):
        if isinstance(v, np.ndarray):
            return float(np.asarray(v).reshape(-1)[0])
        return float(v)

    def _render(self):
        arr = self.env.render()
        return Image.fromarray(np.asarray(arr, dtype=np.uint8), mode="RGB")

    def reset(self):
        self.step_count = 0
        self.env.reset(seed=self.seed)
        return self._render()

    def pause(self):
        pass

    def unpause(self):
        pass

    def close(self):
        self.env.close()

    def _map_gamepad_to_gym_action(self, action):
        lx = self._scalar(action.get("AXIS_LEFTX", 0)) / 32767.0
        if self.discrete_n is not None:
            if self.discrete_n == 2:
                return 1 if lx > 0 else 0
            pressed = [name for name in BUTTON_ACTION_TOKENS if action.get(name, 0) and name in action]
            if pressed:
                return min(BUTTON_ACTION_TOKENS.index(pressed[0]), self.discrete_n - 1)
            return int(np.clip(round((lx + 1.0) * 0.5 * (self.discrete_n - 1)), 0, self.discrete_n - 1))
        # Continuous fallback: repeat left stick into action shape.
        shape = self.env.action_space.shape
        low, high = self.env.action_space.low, self.env.action_space.high
        raw = np.zeros(shape, dtype=np.float32)
        raw.flat[0] = lx
        return np.clip(raw, low, high)

    def step(self, action):
        self.step_count += 1
        gym_action = self._map_gamepad_to_gym_action(action)
        _, reward, terminated, truncated, info = self.env.step(gym_action)
        if terminated or truncated:
            self.env.reset()
        info = dict(info)
        info.update({"debug_env": "gym", "gym_id": self.gym_id, "gym_action": gym_action, "step_count": self.step_count})
        return self._render(), float(reward), bool(terminated), bool(truncated), info


class MockModelClient:
    """Optional model-server replacement for quickly testing script plumbing."""

    def __init__(self, action_len=4, seed=0):
        self.action_len = action_len
        self.rng = np.random.default_rng(seed)

    def reset(self):
        pass

    def info(self):
        return {
            "action_downsample_ratio": 1,
            "ckpt_path": "mock_linux_debug.ckpt",
        }

    def predict(self, obs):
        n = self.action_len
        return {
            "j_left": self.rng.uniform(-1.0, 1.0, size=(n, 2)),
            "j_right": self.rng.uniform(-1.0, 1.0, size=(n, 2)),
            "buttons": self.rng.random(size=(n, len(BUTTON_ACTION_TOKENS))),
        }


class NullVideoRecorder:
    def __init__(self, *args, **kwargs):
        pass

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc, tb):
        return False

    def add_frame(self, frame):
        pass


def parse_args():
    parser = argparse.ArgumentParser(description="VLM Inference")
    parser.add_argument("--process", type=str, default="celeste.exe", help="Game to play")
    parser.add_argument("--allow-menu", action="store_true", help="Allow menu actions (Disabled by default)")
    parser.add_argument("--host", type=str, default="10.245.92.193", help="Host/IP of model server")
    parser.add_argument("--port", type=int, default=5555, help="Port for model server")

    # Linux/debug helpers.
    parser.add_argument(
        "--mock-env",
        action="store_true",
        help="Use the Linux debug environment instead of the real GamepadEnv.",
    )
    parser.add_argument(
        "--debug-env",
        type=str,
        default="gym",
        choices=["visual", "random", "gym"],
        help="Linux/debug env backend. visual is deterministic and action-responsive; random is pure noise; gym uses Gymnasium.",
    )
    parser.add_argument(
        "--gym-id",
        type=str,
        default="CartPole-v1",
        help="Gymnasium environment id used when --debug-env gym.",
    )
    parser.add_argument(
        "--real-env",
        action="store_true",
        help="Force real GamepadEnv even on Linux. Useful only if you really configured GUI/game backend.",
    )
    parser.add_argument(
        "--mock-policy",
        type=bool,
        default=False,
        help="Use random policy outputs instead of connecting to the model server.",
    )
    parser.add_argument(
        "--mock-seed",
        type=int,
        default=0,
        help="Random seed for mock env/policy.",
    )
    parser.add_argument(
        "--mock-width",
        type=int,
        default=1280,
        help="Mock observation width.",
    )
    parser.add_argument(
        "--mock-height",
        type=int,
        default=720,
        help="Mock observation height.",
    )
    parser.add_argument(
        "--mock-action-len",
        type=int,
        default=4,
        help="Number of actions predicted by MockModelClient per step.",
    )
    parser.add_argument(
        "--max-steps",
        type=int,
        default=100,
        help="Stop after N outer-loop steps. 0 means run forever, matching the original script.",
    )
    parser.add_argument(
        "--no-video",
        action="store_true",
        help="Do not write mp4 files. Useful when debugging without ffmpeg/video deps.",
    )
    return parser.parse_args()


def preprocess_img(main_image):
    main_cv = cv2.cvtColor(np.array(main_image), cv2.COLOR_RGB2BGR)
    final_image = cv2.resize(main_cv, (256, 256), interpolation=cv2.INTER_AREA)
    return Image.fromarray(cv2.cvtColor(final_image, cv2.COLOR_BGR2RGB))


def make_zero_action():
    return OrderedDict(
        [
            ("WEST", 0),
            ("SOUTH", 0),
            ("BACK", 0),
            ("DPAD_DOWN", 0),
            ("DPAD_LEFT", 0),
            ("DPAD_RIGHT", 0),
            ("DPAD_UP", 0),
            ("GUIDE", 0),
            ("AXIS_LEFTX", np.array([0], dtype=np.int64)),
            ("AXIS_LEFTY", np.array([0], dtype=np.int64)),
            ("LEFT_SHOULDER", 0),
            ("LEFT_TRIGGER", np.array([0], dtype=np.int64)),
            ("AXIS_RIGHTX", np.array([0], dtype=np.int64)),
            ("AXIS_RIGHTY", np.array([0], dtype=np.int64)),
            ("LEFT_THUMB", 0),
            ("RIGHT_THUMB", 0),
            ("RIGHT_SHOULDER", 0),
            ("RIGHT_TRIGGER", np.array([0], dtype=np.int64)),
            ("START", 0),
            ("EAST", 0),
            ("NORTH", 0),
        ]
    )


def load_policy(args):
    if args.mock_policy:
        policy = MockModelClient(action_len=args.mock_action_len, seed=args.mock_seed)
    else:
        from nitrogen.inference_client import ModelClient

        policy = ModelClient(host=args.host, port=args.port)
        # policy = ModelClient(port=args.port)

    policy.reset()
    policy_info = policy.info()
    return policy, policy_info


def make_env(args, use_mock_env):
    if use_mock_env:
        if args.debug_env == "visual":
            return VisualDebugGamepadEnv(
                game=args.process,
                game_speed=1.0,
                env_fps=60,
                async_mode=True,
                width=args.mock_width,
                height=args.mock_height,
                seed=args.mock_seed,
            )
        if args.debug_env == "gym":
            return GymRGBGamepadEnv(
                game=args.process,
                game_speed=1.0,
                env_fps=60,
                async_mode=True,
                gym_id=args.gym_id,
                seed=args.mock_seed,
            )
        return MockGamepadEnv(
            game=args.process,
            game_speed=1.0,
            env_fps=60,
            async_mode=True,
            width=args.mock_width,
            height=args.mock_height,
            seed=args.mock_seed,
        )

    # Import only when the real environment is requested.
    # On Linux this can fail because game_env imports pyautogui/dxcam.
    from nitrogen.game_env import GamepadEnv

    return GamepadEnv(
        game=args.process,
        game_speed=1.0,
        env_fps=60,
        async_mode=True,
    )


def maybe_init_game_menu(args, env, use_mock_env):
    if use_mock_env:
        return

    # These games require opening a menu to initialize the controller.
    if args.process not in {"isaac-ng.exe", "Cuphead.exe"}:
        return

    print(f"GamepadEnv ready for {args.process} at {env.env_fps} FPS")
    input("Press enter to create a virtual controller and start rollouts...")
    for i in range(3):
        print(f"{3 - i}...")
        time.sleep(1)

    def press(button):
        env.gamepad_emulator.press_button(button)
        env.gamepad_emulator.gamepad.update()
        time.sleep(0.05)
        env.gamepad_emulator.release_button(button)
        env.gamepad_emulator.gamepad.update()

    press("SOUTH")
    for _ in range(5):
        press("EAST")
        time.sleep(0.3)


def main():
    args = parse_args()

    use_mock_env = args.mock_env or (sys.platform != "win32" and not args.real_env)
    if use_mock_env:
        print(f"Using Linux debug env backend: {args.debug_env}")
    else:
        print("Using real GamepadEnv.")

    if args.no_video:
        Recorder = NullVideoRecorder
    else:
        if VideoRecorder is None or create_viz is None:
            raise RuntimeError(
                "Failed to import nitrogen.inference_viz. Re-run with --no-video for lightweight debugging."
            ) from _VIZ_IMPORT_ERROR
        Recorder = VideoRecorder

    policy, policy_info = load_policy(args)
    action_downsample_ratio = int(policy_info.get("action_downsample_ratio", 1))

    ckpt_path = policy_info.get("ckpt_path", "unknown_ckpt")
    CKPT_NAME = Path(ckpt_path).stem
    NO_MENU = not args.allow_menu

    PATH_DEBUG = PATH_REPO / "debug"
    PATH_DEBUG.mkdir(parents=True, exist_ok=True)

    PATH_OUT = (PATH_REPO / "out" / CKPT_NAME).resolve()
    PATH_OUT.mkdir(parents=True, exist_ok=True)

    BUTTON_PRESS_THRES = 0.5
    TOKEN_SET = BUTTON_ACTION_TOKENS

    class RolloutFinished(Exception):
        pass

    video_files = sorted(PATH_OUT.glob("*_DEBUG.mp4"))
    if video_files:
        existing_numbers = [f.name.split("_")[0] for f in video_files]
        existing_numbers = [int(n) for n in existing_numbers if n.isdigit()]
        next_number = max(existing_numbers) + 1 if existing_numbers else 1
    else:
        next_number = 1

    PATH_MP4_DEBUG = PATH_OUT / f"{next_number:04d}_DEBUG.mp4"
    PATH_MP4_CLEAN = PATH_OUT / f"{next_number:04d}_CLEAN.mp4"
    PATH_ACTIONS = PATH_OUT / f"{next_number:04d}_ACTIONS.json"

    zero_action = make_zero_action()

    print("Model loaded, starting environment...")
    if not use_mock_env:
        for i in range(3):
            print(f"{3 - i}...")
            time.sleep(1)

    env = make_env(args, use_mock_env=use_mock_env)
    maybe_init_game_menu(args, env, use_mock_env=use_mock_env)

    env.reset()
    env.pause()

    # Initial call to get state. In debug mode this returns a PIL RGB observation.
    obs, reward, terminated, truncated, info = env.step(action=zero_action)

    step_count = 0

    with Recorder(str(PATH_MP4_DEBUG), fps=60, crf=32, preset="medium") as debug_recorder:
        with Recorder(str(PATH_MP4_CLEAN), fps=60, crf=28, preset="medium") as clean_recorder:
            try:
                while True:
                    obs = preprocess_img(obs)
                    obs.save(PATH_DEBUG / f"{step_count:05d}.png")

                    pred = policy.predict(obs)

                    j_left = np.asarray(pred["j_left"])
                    j_right = np.asarray(pred["j_right"])
                    buttons = np.asarray(pred["buttons"])

                    n = len(buttons)
                    assert n == len(j_left) == len(j_right), "Mismatch in action lengths"

                    env_actions = []

                    for i in range(n):
                        move_action = zero_action.copy()

                        xl, yl = j_left[i]
                        xr, yr = j_right[i]
                        move_action["AXIS_LEFTX"] = np.array([int(xl * 32767)], dtype=np.int64)
                        move_action["AXIS_LEFTY"] = np.array([int(yl * 32767)], dtype=np.int64)
                        move_action["AXIS_RIGHTX"] = np.array([int(xr * 32767)], dtype=np.int64)
                        move_action["AXIS_RIGHTY"] = np.array([int(yr * 32767)], dtype=np.int64)

                        button_vector = buttons[i]
                        assert len(button_vector) == len(TOKEN_SET), (
                            "Button vector length does not match token set length"
                        )

                        for name, value in zip(TOKEN_SET, button_vector):
                            if "TRIGGER" in name:
                                move_action[name] = np.array([int(value * 255)], dtype=np.int64)
                            else:
                                move_action[name] = 1 if value > BUTTON_PRESS_THRES else 0

                        env_actions.append(move_action)

                    print(
                        f"Executing {len(env_actions)} actions, "
                        f"each action will be repeated {action_downsample_ratio} times"
                    )

                    for i, a in enumerate(env_actions):
                        if NO_MENU:
                            if a["START"]:
                                print("Model predicted start, disabling this action")
                            a["GUIDE"] = 0
                            a["START"] = 0
                            a["BACK"] = 0

                        for _ in range(action_downsample_ratio):
                            obs, reward, terminated, truncated, info = env.step(action=a)

                            # if terminated or truncated:
                            #     if terminated:
                            #         print(f"[STOP] terminated=True, balance failed at step={step_count}, substep={i}")
                            #     if truncated:
                            #         print(f"[STOP] truncated=True, episode time limit reached at step={step_count}, substep={i}")

                            #     raise RolloutFinished

                            if not args.no_video:
                                obs_viz = np.array(obs).copy()
                                clean_viz = cv2.resize(
                                    obs_viz,
                                    (1920, 1080),
                                    interpolation=cv2.INTER_AREA,
                                )
                                debug_viz = create_viz(
                                    cv2.resize(
                                        obs_viz,
                                        (1280, 720),
                                        interpolation=cv2.INTER_AREA,
                                    ),
                                    i,
                                    j_left,
                                    j_right,
                                    buttons,
                                    token_set=TOKEN_SET,
                                )
                                debug_recorder.add_frame(debug_viz)
                                clean_recorder.add_frame(clean_viz)

                    with open(PATH_ACTIONS, "a") as f:
                        for i, a in enumerate(env_actions):
                            serializable_action = {}
                            for k, v in a.items():
                                if isinstance(v, np.ndarray):
                                    serializable_action[k] = v.tolist()
                                else:
                                    serializable_action[k] = v
                            serializable_action["step"] = step_count
                            serializable_action["substep"] = i
                            json.dump(serializable_action, f)
                            f.write("\n")

                    step_count += 1
                    if args.max_steps > 0 and step_count >= args.max_steps:
                        print(f"Reached --max-steps={args.max_steps}, exiting.")
                        break
            
            except RolloutFinished:
                print("Rollout finished. Closing environment...")
            
            finally:
                env.unpause()
                env.close()


if __name__ == "__main__":
    main()